Dating birth-related clavicular fractures: pediatric radiologists versus artificial intelligence

BackgroundFracture dating from skeletal surveys is crucial in the diagnosis and investigation of infant abuse. However, this task is challenging because of the subjective nature of the radiologic interpretation and the lack of ground truth. Researchers have used birth-related clavicle fractures as a surrogate to study the radiographic pattern of healing; however, they did not elucidate the accuracy performance of the radiologists in dating fractures.ObjectiveTo determine the accuracy of radiologists in dating birth-related clavicle fractures and compare their performance to that achieved by computer algorithm.Materials and methodsWe used a previously assembled birth-related clavicle fracture database consisting of 416 anteroposterior clavicle radiographs as the study cohort. The average and standard deviation of the fracture age within this database were 24 days and 18 days, respectively. Three blinded radiologists independently estimated the ages of the clavicle fractures depicted in the radiographs within the database. We compared these estimation results to those made by a recently published deep-learning (DL) model conducted with the identical infant cohort. We calculated standard error metrics to compare the accuracy performances of the radiologists and the computer model.ResultsThe intra- and inter-reader agreements of the fracture age estimates by the radiologists were moderate to good. The radiologists estimated the fracture ages with a mean absolute error (MAE) of 6.1-7.1 days, and standard deviation of the absolute error of 6.3-8.3 days. The accuracy performances of the three radiologists were not significantly different from one another. In comparison, the DL model estimated the age of clavicle fractures with an MAE of 4.2 days, significantly lower than all of the radiologists (P < 0.001).ConclusionThree experienced pediatric radiologists dated clavicular fractures with moderate-good intra- and inter-reader agreements. The correlations between the radiologists' estimates and the ground truth were moderate to good. The fracture ages assigned by the DL model showed superior correlation with the ground truth compared to radiologists' dating estimates.

Automated summary

This study aimed to determine the accuracy of radiologists in dating birth-related clavicle fractures and compare their performance to that achieved by a computer algorithm. The results showed that three experienced pediatric radiologists had moderate to good intra- and inter-reader agreements in dating clavicle fractures. Compared to the radiologists, the deep-learning (DL) model had a lower mean absolute error and showed better correlation with the ground truth.